PyTorch
定义
PyTorch is a popular deep learning framework with an imperative, Pythonic style. 它是 widely used in research and industry and is the base for many libraries (Hugging Face, torchvision, etc.).
It competes with TensorFlow; PyTorch is often preferred for research and rapid iteration due to eager execution and clear debugging. Used for LLMs, vision, NLP, and RL; production deployment uses TorchScript, ONNX, or dedicated runtimes.
工作原理
你将模型定义为 Python 类（nn.Module）；前向传播立即运行（eager 模式），autograd records operations for backprop. Optimizers (例如 Adam) update parameters from computed gradients. Training loops are explicit (for epoch, for batch, loss.backward(), optimizer.step()), which makes experimentation and debugging straightforward. For deployment, export to TorchScript (Python-free), ONNX (cross-framework), or PyTorch Mobile. Distributed training uses torch.distributed or higher-level APIs (例如 Hugging Face Accelerate). The same code can run on CPU, GPU, or multiple GPUs with minimal changes.
应用场景
PyTorch is a natural fit for research, prototyping, and production when you want flexibility and a Python-first workflow.
- Research and prototyping (例如 new architectures, experiments)
- Training and fine-tuning models (vision, NLP, RL)
- Production deployment via TorchScript, ONNX, or mobile